1 Introduction

Here, we will apply a k-nearest neighbor (KNN) algorithm to classify the scATAC cells to a given cell type category with the help of our training set, the Multiome experiment. Remember, that KNN works on a basic assumption that data points of similar categories are closer to each other.

2 Pre-processing

2.1 Load packages

library(Seurat)
library(Signac)
library(flexclust)
library(tidyverse)
library(plyr)
library(harmony)
library(class)
library(ggplot2)
library(reshape2)
library(ggpubr)

set.seed(222)

2.2 Parameters

cell_type = "CD4_T"

# Paths
path_to_obj <- str_c(
  here::here("scATAC-seq/results/R_objects/level_5/"),
  cell_type,
  "/01.",
  cell_type,
  "_integrated_level_5.rds",
  sep = ""
)

path_to_obj_RNA <- str_c(
  here::here("scRNA-seq/3-clustering/5-level_5/"),
  cell_type,
    "/",
  cell_type,
  "_subseted_annotated_level_5.rds",
  sep = ""
)

path_to_level_4 <- here::here("scATAC-seq/results/R_objects/level_5/CD4_T/")
path_to_save <- str_c(path_to_level_4, "02.CD4_T_annotated_level_5.rds")

2.3 Variables

reduction <- "harmony"
dims <- 1:40

color_palette <-  c("#1CFFCE", "#90AD1C", "#C075A6", "#85660D", 
                    "#5A5156", "#AA0DFE", "#F8A19F", "#F7E1A0",
                    "#1C8356", "#FEAF16", "#822E1C", "#C4451C", 
                    "#1CBE4F", "#325A9B", "#F6222E", "#FE00FA",
                    "#FBE426", "#16FF32",  "black",  "#3283FE",
                    "#B00068", "#DEA0FD", "#B10DA1", "#E4E1E3", 
                    "#90AD1C", "#FE00FA", "#85660D", "#3B00FB", 
                    "#822E1C", "coral2",  "#1CFFCE", "#1CBE4F", 
                    "#3283FE", "#FBE426", "#F7E1A0", "#325A9B", 
                    "#2ED9FF", "#B5EFB5", "#5A5156", "#DEA0FD",
                    "#FEAF16", "#683B79", "#B10DA1", "#1C7F93", 
                    "#F8A19F", "dark orange", "#FEAF16", 
                    "#FBE426", "Brown")

2.4 Load data

We need to load the scRNAseq annotation from Multiome experiment (cell barcode and cell-type assigned) and the integrated scATAC data.

seurat <- readRDS(path_to_obj_RNA)

tonsil_RNA_annotation <- seurat@meta.data %>%
  rownames_to_column(var = "cell_barcode") %>%
  dplyr::filter(assay == "multiome") %>%
  dplyr::select("cell_barcode", "annotation_paper")
head(tonsil_RNA_annotation)
##                           cell_barcode annotation_paper
## 1 co7dzuup_xuczw9vc_AAACATGCAAGCCAGA-1      GC-Tfh-0X40
## 2 co7dzuup_xuczw9vc_AAACATGCAAGGTATA-1            Naive
## 3 co7dzuup_xuczw9vc_AAACCGGCATGCTATG-1        Tfh-LZ-GC
## 4 co7dzuup_xuczw9vc_AAACGCGCATTGTGTG-1            Naive
## 5 co7dzuup_xuczw9vc_AAAGCGGGTTTGGGCG-1       GC-Tfh-SAP
## 6 co7dzuup_xuczw9vc_AAATGCCTCACCTGTC-1      GC-Tfh-0X40
DimPlot(seurat,
  group.by = "annotation_paper",
  cols = color_palette,
  pt.size = 0.1)

seurat_ATAC <- readRDS(path_to_obj)
seurat_ATAC
## An object of class Seurat 
## 93602 features across 16383 samples within 1 assay 
## Active assay: peaks_redefined (93602 features, 93293 variable features)
##  3 dimensional reductions calculated: umap, lsi, harmony
p1 <- DimPlot(seurat_ATAC,
  pt.size = 0.1)
p1

Annotation level 5 for scATAC will be defined “a priori” as unannotated and the scRNA annotation will be transfered to the scATAC-multiome cells based on the same cell barcode.

tonsil_scATAC_df <- data.frame(cell_barcode = colnames(seurat_ATAC)[seurat_ATAC$assay == "scATAC"])
tonsil_scATAC_df$annotation_paper <- "unannotated"

df_all <- rbind(tonsil_RNA_annotation,tonsil_scATAC_df)
rownames(df_all) <- df_all$cell_barcode
df_all <- df_all[colnames(seurat_ATAC), ]

seurat_ATAC$annotation_paper <- df_all$annotation_paper
seurat_ATAC@meta.data$annotation_prob  <- 1
melt(table(seurat_ATAC$annotation_paper))
##              Var1 value
## 1           Naive  2551
## 2  CM Pre-non-Tfh  1233
## 3       CM PreTfh   276
## 4     T-Trans-Mem   233
## 5       T-Eff-Mem   306
## 6        T-helper   340
## 7  Tfh T:B border    26
## 8       Tfh-LZ-GC  1299
## 9      GC-Tfh-SAP  1176
## 10    GC-Tfh-0X40   238
## 11        Tfh-Mem   467
## 12      Eff-Tregs   477
## 13 non-GC-Tf-regs   155
## 14     GC-Tf-regs   240
## 15    unannotated  7366
table(is.na(seurat_ATAC$annotation_paper))
## 
## FALSE 
## 16383

2.5 General low-dimensionality representation of the assays

DimPlot(seurat_ATAC,
  group.by = "annotation_paper",
  split.by = "assay",
  cols = color_palette,
  pt.size = 0.5)

3 KNN Algorithm

3.1 Data Splicing

Data splicing basically involves splitting the data set into training and testing data set.

reference_cells <- colnames(seurat_ATAC)[seurat_ATAC$assay == "multiome"]
query_cells <- colnames(seurat_ATAC)[seurat_ATAC$assay == "scATAC"]

reduction_ref <- seurat_ATAC@reductions[[reduction]]@cell.embeddings[reference_cells, dims]
reduction_query <- seurat_ATAC@reductions[[reduction]]@cell.embeddings[query_cells, dims]

3.2 Cross-validation of the K parameter.

We’re going to calculate the number of observations in the training dataset that correspond to the Multiome data. The reason we’re doing this is that we want to initialize the value of ‘K’ in the KNN model. To do that, we split our training data in two part: a train.loan, that correspond to the random selection of the 70% of the training data and the test.loan, that is the remaining 30% of the data set. The first one is used to traint the system while the second is uses to evaluate the learned system.

dat.d <- sample(1:nrow(reduction_ref),
               size=nrow(reduction_ref)*0.7,replace = FALSE) 

train.loan  <- reduction_ref[dat.d,] # 70% training data
test.loan <- reduction_ref[-dat.d,] # remaining 30% test data

train.loan_labels <- seurat_ATAC@meta.data[row.names(train.loan),]$annotation_paper
test.loan_labels <- seurat_ATAC@meta.data[row.names(test.loan),]$annotation_paper

k.optm <- c()
k.values <- c()

for (i in c(2,4,8,10,12,14,16,32,64)){
 print(i)
 knn.mod <- knn(train=train.loan, test=test.loan, cl=train.loan_labels, k=i)
 k.optm <- c(k.optm, 100 * sum(test.loan_labels == knn.mod)/NROW(test.loan_labels))
 k.values <- c(k.values,i)
}
## [1] 2
## [1] 4
## [1] 8
## [1] 10
## [1] 12
## [1] 14
## [1] 16
## [1] 32
## [1] 64

Now we can plot the accuracy of the model taking in account a range of different K and selec the best one.

k.optim = data.frame(k.values,k.optm)

p3 <- ggplot(data=k.optim, aes(x=k.values, y=k.optm, group=1)) +
 geom_line() +
 geom_point() + 
 geom_vline(xintercept=10, linetype="dashed", color = "red")

p3

3.3 Building a Machine Learning model with the optimal k value.

train.loan  <- reduction_ref
test.loan <- reduction_query

train.loan_labels <- seurat_ATAC@meta.data[row.names(train.loan),]$annotation_paper
test.loan_labels <- seurat_ATAC@meta.data[row.names(test.loan),]$annotation_paper

knn.mod <- knn(train=train.loan, test=test.loan, cl=train.loan_labels, k=10, prob=T)

annotation_data <- data.frame(query_cells, knn.mod, attr(knn.mod,"prob"))
colnames(annotation_data) <- c("query_cells",
                               "annotation_paper",
                               "annotation_prob")

annotation_data$annotation_paper <- as.character(annotation_data$annotation_paper)
seurat_ATAC@meta.data[annotation_data$query_cells,]$annotation_paper <- annotation_data$annotation_paper
seurat_ATAC@meta.data[annotation_data$query_cells,]$annotation_prob <- annotation_data$annotation_prob
seurat_ATAC$annotation_paper <- factor(seurat_ATAC$annotation_paper)

3.4 Low-dimensionality representation of the assays

DimPlot(
  seurat_ATAC,
  cols = color_palette,
  group.by = "annotation_paper",
  pt.size = 0.1) + ggtitle("")

#ggsave(path = "/Users/pauli/Desktop/", 
 #      filename = "CD4_T_umap_level_5_ATAC.png",
  #     width = 20, height = 15, units = "cm")

DimPlot(
  cols = color_palette,
  seurat_ATAC, reduction = "umap",
  group.by = "annotation_paper",
  pt.size = 0.1,  split.by = "assay")

saveRDS(seurat_ATAC, path_to_save)

3.4.1 Percentage of cells per CD4-T subtype

general_counts_melt <-  melt(table(seurat_ATAC$annotation_paper))

ggdotchart(general_counts_melt, 
     x = "Var1", 
     y = "value",
     xlab = FALSE,
     ylab = FALSE,
     sorting    = "none",
     add = "segments",    
     color = 'gray80',
     rotate = TRUE,                             
     dot.size = 10, 
     label = round(general_counts_melt$value),   
     font.label = list(color = "black", 
                       size = 9, vjust = 0.5), 
     ggtheme = theme_pubr()) + 
    theme(axis.text.x = element_text(angle = 90, 
                                     hjust = 1, size=9),
    axis.text.y = element_text(size=9)) + 
    scale_y_continuous(limits=c(0, 5000)) 

3.5 Low-dimensionality representation of the prediction probability

Note that the probability of the prediction was lower in the transitioning cells and in not-defined clusters.

seurat_ATAC_scATAC = subset(seurat_ATAC, assay == "scATAC")

FeaturePlot(
  seurat_ATAC_scATAC, reduction = "umap",
  features = "annotation_prob",
  pt.size = 0.1)

4 Session Information

sessionInfo()
## R version 4.0.3 (2020-10-10)
## Platform: x86_64-apple-darwin13.4.0 (64-bit)
## Running under: macOS Big Sur 10.16
## 
## Matrix products: default
## BLAS/LAPACK: /Users/pauli/opt/anaconda3/envs/Motif_TF/lib/libopenblasp-r0.3.10.dylib
## 
## locale:
## [1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
## 
## attached base packages:
## [1] stats4    grid      stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
##  [1] ggpubr_0.4.0       reshape2_1.4.4     class_7.3-17       harmony_1.0        Rcpp_1.0.6         plyr_1.8.6         forcats_0.5.0      stringr_1.4.0      dplyr_1.0.7        purrr_0.3.4        readr_1.4.0        tidyr_1.1.3        tibble_3.1.2       ggplot2_3.3.5      tidyverse_1.3.0    flexclust_1.4-0    modeltools_0.2-23  lattice_0.20-41    Signac_1.2.1       SeuratObject_4.0.2 Seurat_4.0.3       BiocStyle_2.16.1  
## 
## loaded via a namespace (and not attached):
##   [1] utf8_1.2.1             reticulate_1.20        tidyselect_1.1.1       htmlwidgets_1.5.3      docopt_0.7.1           BiocParallel_1.22.0    Rtsne_0.15             munsell_0.5.0          codetools_0.2-17       ica_1.0-2              future_1.21.0          miniUI_0.1.1.1         withr_2.4.2            colorspace_2.0-2       knitr_1.30             rstudioapi_0.11        ROCR_1.0-11            ggsignif_0.6.0         tensor_1.5             listenv_0.8.0          labeling_0.4.2         slam_0.1-47            GenomeInfoDbData_1.2.3 polyclip_1.10-0        farver_2.1.0           rprojroot_2.0.2        parallelly_1.26.1      vctrs_0.3.8            generics_0.1.0         xfun_0.18              lsa_0.73.2             ggseqlogo_0.1          R6_2.5.0               GenomeInfoDb_1.24.0    bitops_1.0-7           spatstat.utils_2.2-0   assertthat_0.2.1       promises_1.2.0.1       scales_1.1.1           gtable_0.3.0           globals_0.14.0         goftest_1.2-2          rlang_0.4.11           RcppRoll_0.3.0         splines_4.0.3          rstatix_0.6.0          lazyeval_0.2.2         spatstat.geom_2.2-0    broom_0.7.2            BiocManager_1.30.10    yaml_2.2.1            
##  [52] abind_1.4-5            modelr_0.1.8           backports_1.1.10       httpuv_1.6.1           tools_4.0.3            bookdown_0.21          ellipsis_0.3.2         spatstat.core_2.2-0    RColorBrewer_1.1-2     BiocGenerics_0.34.0    ggridges_0.5.3         zlibbioc_1.34.0        RCurl_1.98-1.2         rpart_4.1-15           deldir_0.2-10          pbapply_1.4-3          cowplot_1.1.1          S4Vectors_0.26.0       zoo_1.8-9              haven_2.3.1            ggrepel_0.9.1          cluster_2.1.0          here_1.0.1             fs_1.5.0               magrittr_2.0.1         data.table_1.14.0      scattermore_0.7        openxlsx_4.2.3         lmtest_0.9-38          reprex_0.3.0           RANN_2.6.1             SnowballC_0.7.0        fitdistrplus_1.1-5     matrixStats_0.59.0     hms_0.5.3              patchwork_1.1.1        mime_0.11              evaluate_0.14          xtable_1.8-4           rio_0.5.16             sparsesvd_0.2          readxl_1.3.1           IRanges_2.22.1         gridExtra_2.3          compiler_4.0.3         KernSmooth_2.23-17     crayon_1.4.1           htmltools_0.5.1.1      mgcv_1.8-33            later_1.2.0            lubridate_1.7.9       
## [103] DBI_1.1.0              tweenr_1.0.1           dbplyr_1.4.4           MASS_7.3-53            Matrix_1.3-4           car_3.0-10             cli_3.0.0              parallel_4.0.3         igraph_1.2.6           GenomicRanges_1.40.0   pkgconfig_2.0.3        foreign_0.8-80         plotly_4.9.4.1         spatstat.sparse_2.0-0  xml2_1.3.2             XVector_0.28.0         rvest_0.3.6            digest_0.6.27          sctransform_0.3.2      RcppAnnoy_0.0.18       spatstat.data_2.1-0    Biostrings_2.56.0      rmarkdown_2.5          cellranger_1.1.0       leiden_0.3.8           fastmatch_1.1-0        uwot_0.1.10            curl_4.3.2             shiny_1.6.0            Rsamtools_2.4.0        lifecycle_1.0.0        nlme_3.1-150           jsonlite_1.7.2         carData_3.0-4          viridisLite_0.4.0      fansi_0.5.0            pillar_1.6.1           fastmap_1.1.0          httr_1.4.2             survival_3.2-7         glue_1.4.2             zip_2.1.1              qlcMatrix_0.9.7        png_0.1-7              ggforce_0.3.2          stringi_1.6.2          blob_1.2.1             irlba_2.3.3            future.apply_1.7.0